Upload 2 files

README.md

@@ -3,38 +3,48 @@
 This directory contains utilities for the purpose of browsing the
 "token space" of CLIP ViT-L/14
 
-Long term goal is to be able to literally browse from word to 
-"nearby" word
+Primary tool is "generate-distances.py", 
+which allows command-line browsing of words and their neighbours
+
+
+## generate-distances.py
+
+Loads the generated embeddings, calculates a full matrix
+of distances between all tokens, and then reads in a word, to show neighbours for.
+
+To run this requires the files "embeddings.safetensors" and "fullword.json"
 
 
 
 ## generate-embeddings.py
 
-Generates the "embeddings.safetensor" file!
+Generates the "embeddings.safetensor" file. Takes a few minutes to run.
 
 Basically goes through the fullword.json file, and 
-generates a standalone embedding object for each word.
+generates a standalone embedding for each word.
 Shape of the embeddings tensor, is
  [number-of-words][768]
 
-Note that it is possible to directly pull a tensor from the CLIP model,
-key of  text_model.embeddings.token_embedding.weight
+Note that yes, it is possible to directly pull a tensor from the CLIP model,
+using keyname of  text_model.embeddings.token_embedding.weight
 
 This will NOT GIVE YOU THE RIGHT DISTANCES!
 Hence why we are calculating and then storing the embedding weights actually
-used by the CLIP process
+generated by the CLIP process
 
+## embeddings.safetensors
 
-## generate-distances.py
+Data file generated by generate-embeddings.py
 
-Loads the prior generated embeddings, and then tries to calculate a full matrix
-of distances between all tokens
 
 
 ## fullword.json
 
 This file contains a collection of "one word, one CLIP token id" pairings.
-The file was taken from vocab.sdturbo.json
+The file was taken from vocab.json, which is part of multiple SD models in huggingface.co
+
+The file was optimized for what people are actually going to type as words.
 First all the non-(/w) entries were stripped out.
 Then all the garbage punctuation and foreign characters were stripped out.
 Finally, the actual (/w) was stripped out, for ease of use.
+

generate-distances.py

@@ -23,44 +23,40 @@ with open("fullword.json","r") as f:
     tokendict = json.load(f)
 wordlist = list(tokendict.keys())
 
-print("read in embeddingsnow",file=sys.stderr)
-
+print("read in embeddings now",file=sys.stderr)
 
 model = safe_open(embed_file,framework="pt",device="cuda")
 embs=model.get_tensor("embeddings")
 embs.to(device)
-
-
 print("Shape of loaded embeds =",embs.shape)
-print("calculate distances now")
 
+# ("calculate distances now")
 distances = torch.cdist(embs, embs, p=2)
 print("distances shape is",distances.shape)
 
-targetword="cat"
-targetindex=wordlist.index(targetword)
-print("index of cat is",targetindex)
-targetdistances=distances[targetindex]
-
-smallest_distances, smallest_indices = torch.topk(targetdistances, 5, largest=False)
+# Find 10 closest tokens to targetword.
+# Will include the word itself
+def find_closest(targetword):
+    try:
+        targetindex=wordlist.index(targetword)
+    except ValueError:
+        print(targetword,"not found")
+        return
 
-smallest_distances=smallest_distances.tolist()
-smallest_indices=smallest_indices.tolist()
+    #print("index of",targetword,"is",targetindex)
+    targetdistances=distances[targetindex]
 
-print("The smallest distance values are",smallest_distances)
-print("The smallest index values are",smallest_indices)
+    smallest_distances, smallest_indices = torch.topk(targetdistances, 10, largest=False)
 
-for t in smallest_indices:
-    print(wordlist[t])
+    smallest_distances=smallest_distances.tolist()
+    smallest_indices=smallest_indices.tolist()
+    for d,i in zip(smallest_distances,smallest_indices):
+        print(wordlist[i],"(",d,")")
+    #print("The smallest distance values are",smallest_distances)
+    #print("The smallest index values are",smallest_indices)
 
 
-
-"""
-import torch.nn.functional as F
-pos=0
-for word in tokendict.keys():
-    print("Calculating distances from",word)
-    home=embs[pos]
-    #distances = torch.cdist(embs, home.unsqueeze(0), p=2)
-    #distance = F.pairwise_distance(home, embs[,p=2).item()
-"""
+print("Input a word now:")
+for line in sys.stdin:
+    input_text = line.rstrip()
+    find_closest(input_text)